For Judges and Reviewers

This is the current 2026-05-25 verification path for the DueCare Gemma 4 Good Hackathon submission. Older notebook-era material is archived or marked historical; the active Kaggle path is exactly the three script kernels listed below.

Thirty-Second Summary

DueCare is a Gemma 4 harness ecosystem for migrant-worker exploitation risk. It wraps Gemma 4 with reusable safety layers, knowledge packs, deterministic tools, sensitive-data handling, evaluation, and report generation. The current proof focuses on three active surfaces:

Surface	Purpose
kaggle/01-duecare-exploration-workbench/	Broad interactive workbench: chat, harness comparison, search, knowledge extraction, bulk review, traces, and activity logs.
kaggle/02-live-demo/	Focused live demo and video path.
kaggle/A-00-omni-experiment-workbench/	Quantitative proof path: baseline, harnessed, synthetic-data, fine-tuning, judging, checkpoints, and report bundles.

The active inventory is tracked in kaggle/_INDEX.md and docs/current_kaggle_notebook_state.md. A-00 is active as the proof and training/evaluation kernel. It is not required for the shortest recording path, but it should remain runnable and copyable to Kaggle while judging is active.

What To Verify First

1. Open 01-duecare-exploration-workbench and run the chat/harness comparison path. Confirm the default harness uses Persona + GREP + RAG/context + tools, with online search/imports only when explicitly enabled.
2. Open A-00-omni-experiment-workbench and run the preconfigured path for a small prompt count if new proof artifacts are needed. Confirm the Activity log shows the numbered pipeline steps, prompt/response artifacts, combined rule + LLM grading, and report bundle links under /kaggle/working.
3. Confirm all local inference model loading goes through Gemma4Runtime.load() and the Unsloth FastModel recipe documented in docs/model_loading_trace.md.

Current Harness Contract

The authoritative harness docs are:

• docs/harness_ecosystem.md - registered harness inventory and broader harness families.
• docs/harness_pattern.md - module contract and active Kaggle integration pattern.
• docs/harness_standard_contract.md - universal HarnessSpec, model targets, transports, and pack contracts.

The registered harnesses are chat, process, extraction, anonymization, search_safety, post_search_verification, search, and import_corpus. A-00 also uses pipeline-specific harness families for synthetic data generation, fine-tuning, combined judging, checkpointing, and report/export bundles.

A-00 Proof Path

The preconfigured active A-00 pipeline should:

1. Check/unload current model state.
2. Check disk space and clean if needed.
3. Load the selected Gemma model through the shared runtime.
4. Run prompts without the harness.
5. Run the same prompts with the offline DueCare harness.
6. Generate synthetic SFT rows from harnessed outputs.
7. Optionally fine-tune a LoRA adapter with checkpoint/resume enabled.
8. Run fine-tuned no-harness and fine-tuned harnessed arms when training is enabled.
9. Load the selected judge model or configured external judge.
10. Grade all response sets with combined rule + LLM judging.
11. Write HTML, Markdown, JSON, CSV, SVG, PDF-summary, activity, and evidence ZIP artifacts.

Default offline harness behavior should match the Kernel 01 comparison path: Persona + GREP + RAG/context + deterministic tools, with internet and import off for the default proof run.

Verification Commands

From the repository root:

$env:PYTHONPATH='packages/duecare-llm-models/src;packages/duecare-llm-chat/src;packages/duecare-llm-core/src'

python -m py_compile `
  kaggle\A-00-omni-experiment-workbench\kernel.py `
  scripts\generate_notebook_guide.py `
  scripts\kaggle_notebook_utils.py

python -m pytest `
  tests\test_a00_runtime_and_parity_contract.py `
  tests\test_a00_notebook_contract.py `
  tests\test_harness_universal_model_contract.py `
  tests\test_harness_standard_contract.py `
  tests\test_harness_imports.py `
  tests\test_harness_ecosystem_docs.py `
  tests\test_kaggle_notebook_utils.py `
  packages\duecare-llm-chat\tests\test_harness_workbench.py `
  packages\duecare-llm-chat\tests\test_workbench_inventory_integrity.py `
  packages\duecare-llm-models\tests\test_models_package_smoke.py `
  -q --basetemp=.pytest_tmp

What This Submission Does Not Claim

• It does not require archived notebook-era surfaces for the current judge path.
• It does not require paid external judge APIs; Anthropic and Ollama judge routes are optional.
• It does not send raw private case material to a public search engine in the default proof path.
• It does not treat generated or imported knowledge as automatic truth; human review and provenance remain part of the workflow.